Harper Reed's LLM Codegen Workflow

Table of content
Harper Reed's LLM Codegen Workflow

Harper Reed is a Chicago-based technologist who was CTO of Threadless from 2005-2009, then CTO of Obama’s 2012 re-election campaign where he led a 40-person engineering team that raised over $700 million online. After Obama, he founded Modest (acquired by PayPal), where he became Head of Commerce at Braintree.

His February 2025 blog post on LLM code generation became a widely-shared reference for structuring AI-assisted development.

The Three-Phase Workflow

Reed’s tl;dr:

“Brainstorm spec, then plan a plan, then execute using LLM codegen. Discrete loops. Then magic.”

PhaseToolOutput
BrainstormGPT-4oRefined idea, clear requirements
PlanBest reasoning modelDetailed spec and architecture
ExecuteClaude Code / AiderWorking code

The key insight: discrete loops. Don’t let phases bleed together. See The Three-Layer Workflow for how this maps to a broader pattern.

Greenfield vs Legacy

Reed handles two scenarios:

Greenfield code:

  1. Chat to hone the idea
  2. Generate spec with reasoning model
  3. Execute with agentic tool

Legacy modern code:

  1. Feed existing codebase context
  2. Plan modifications
  3. Execute incrementally

Both follow the same brainstorm → plan → execute structure.

The Hero’s Journey

Reed’s follow-up post (April 2025) addresses beginners:

“If I were starting out, I don’t know if it is helpful to jump right into the ‘agent’ coders. It is annoying and weird.”

Recommended progression:

  1. Copilot — Autocomplete, low commitment
  2. Copy-paste from Claude web — Manual but educational
  3. Cursor/Continue — Editor integration
  4. Full agents — Claude Code, Aider

Each stage builds intuition for the next.

Basic Claude Code

After Claude Code’s release, Reed migrated from Aider (May 2025):

“Claude Code was released eight days after I wrote my original workflow blog post, and as I predicted, it made a lot of my post irrelevant.”

His updated workflow:

  1. Chat with GPT-4o to hone idea
  2. Best reasoning model for spec generation
  3. Claude Code for execution

Claude Code is “powerful, and a hell of a lot more expensive.”

The Dr. Biz Persona

In a TWiT interview (September 2025), Reed revealed his quirky habit of assigning AI nicknames. He calls himself “Dr. Biz” when prompting Claude Code — personality-driven prompting that he finds makes interactions more productive.

Waterfall in 15 Minutes

Reed’s April 2025 post challenges traditional software craft:

“For years, we’ve talked about code as craft — how we get into that precious flow state, sculpt a piece of logic, and emerge victorious. But there’s a new paradigm creeping in…”

With AI-assisted development, the question becomes: why does code quality matter the same way when you can regenerate it?

Background

RolePeriod
CTO, Threadless2005-2009
CTO, Obama 2012Campaign technology
Founder/CEO, Modest2013-2015 (acquired by PayPal)
Head of Commerce, Braintree/PayPal2015-2018
CEO, General Galactic2020-2023 (web3/NFT)
CEO, 2389 ResearchCurrent

Reed builds “many small products using LLMs” and documents his evolving practices on his blog and GitHub.

Community Influence

Reed’s original post sparked responses:

Key Insights

PrincipleApplication
Discrete phasesDon’t mix brainstorm, plan, execute
Use the right modelChat for ideas, reasoning for specs, agents for code
Progress through stagesDon’t start with full agents
Document what works nowPractices change rapidly
Personality in prompts“Dr. Biz” makes it more fun

Next: Steve Krouse’s Val Town Platform

Topics: ai-coding workflow prompting automation